English · 中文
Multi-voice deliberation + calibration discipline + spatial audit for regulated AI workflows.
A research monorepo producing the open-source AI infrastructure for regulated decision domains — trading book changes, loan origination, wealth-advisory recommendations, and other workflows where confident-but-wrong agents leave no audit trail. Each decision is debated by five disagreement-by-design AI voices, adjudicated, Brier-audited over time, and surfaced through a spatial extended-reality scene with a tamper-evident hash-chain audit trail.
| Trading vertical Macro · Sector · Portfolio · Growth-VC · Activist Short  |
Banking vertical Credit Fundamentals · Risk Officer · Fair Lending · Customer Advocate · Macro Contrarian  |
Status: active research codebase. Capstone project (trading vertical) advised by Prof. Michael Yang. Banking vertical with Dr. NGO and external risk-math collaborator Loredana C. Levitchi. Submission targets: ICAIF 2027 (trading), IEEE journal (banking).
Adversarial robustness: see docs/threat-model.md for the per-attack catalog + mitigation status. First entry (MosaicLeaks multi-turn information leak, ServiceNow+HuggingFace 2026-06) filed 2026-06-21 via issue #1.
License: MIT.
Production AI agents in regulated industries fail not because they are slow, but because they are confident-but-wrong, leave no audit trail, and have no per-voice calibration history. Single-voice agentic systems propagate the foundation model's confidence without measuring whether the model has been right on this kind of decision before.
The Embodied Compliance Council addresses three gaps directly:
- Epistemic diversity over single-model confidence. Five voices restricted to distinct evidence subsets (factor exposures, sector dynamics, fair-lending statistics, etc.) so the council does not collapse to the foundation model's default opinion.
- Calibration as the audit primitive. Every voice's verdict is logged with confidence and Brier-scored against outcomes. Voices whose dissents proved load-bearing receive higher weight; voices whose certainty was noise receive lower weight.
- Spatial, tamper-evident audit. A WebXR scene renders the council as physical podiums around the user, and every adjudication chains forward via WebCrypto SHA-256 — meeting both EU AI Act Article 14 (human oversight) and U.S. ECOA / Reg B (auditable decisioning) on the same primitive.
embodied-compliance-council/
├── packages/
│ ├── council-voices/ # 10 system-prompt voices (Markdown + YAML)
│ │ ├── trading/ # 5 trading voices (capstone vertical)
│ │ │ ├── 01-macro.md
│ │ │ ├── 02-sector.md
│ │ │ ├── 03-portfolio.md
│ │ │ ├── 04-growth-vc.md
│ │ │ └── 05-activist-short.md
│ │ └── banking/ # 5 banking voices (Dr. NGO vertical)
│ │ ├── 01-credit-fundamentals.md
│ │ ├── 02-risk-officer.md
│ │ ├── 03-fair-lending.md
│ │ ├── 04-customer-advocate.md
│ │ └── 05-macro-contrarian.md
│ ├── orallexa-risk/ # 7 risk primitives (Python, Phase A scaffold)
│ ├── spatial-gating-protocol/ # Shared TypeScript types + WebCrypto hash chain
│ └── hand-gestures/ # MediaPipe Hands gesture classifier (Phase 2.5)
├── apps/
│ ├── trading-quest/ # WebXR scene for the trading vertical
│ └── banking-quest/ # WebXR scene for the banking vertical
├── docs/
│ ├── capstone-thesis/ # Capstone (Prof. Yang) amendments + materials
│ ├── dr-ngo-final/ # Banking proposal v2 + Lora handoff
│ ├── katz-rfp-2026/ # Katz Faculty Research Initiative submission packet
│ ├── ibm-bob-spike/ # IBM AI Builders Challenge weekend spike guide
│ ├── checkpoints/ # Cross-session recovery snapshots
│ └── startup/ # Bottom-up TAM + disruption framework (academic defense)
└── README.md
Each voice has YAML frontmatter (voice_id, allowed_tools, verdict_options, weight_in_aggregate_default) plus a system prompt with a strict-JSON output schema. Tool restrictions are what give the council epistemic diversity — if everyone could see everything, they would converge to the foundation model's default opinion.
| Voice | Lens | Allowed tools |
|---|---|---|
| Macro Analyst | Factor regimes + correlation history | factor_exposures, correlation_matrix, sector_exposure |
| Sector Specialist | GICS sub-industry catalysts + co-movement | beta_decomposition, sector_exposure |
| Portfolio Manager | NAV concentration + VaR step changes | concentration, var, correlation_matrix |
| Growth-VC Contrarian | 3-year moat + value/growth tilt drift | factor_exposures, beta_decomposition |
| Activist Short | Crowded-trade detection + asymmetric break | concentration, sector_exposure |
| Voice | Lens | Allowed tools |
|---|---|---|
| Credit Fundamentals Analyst | Buffett-style through-cycle ability to repay | cash_flow_coverage, debt_service_history, through_cycle_earnings |
| Risk Officer | PD × LGD × EAD + stress scenarios | default_probability, expected_loss, stress_scenario |
| Fair Lending Reviewer | ECOA / Reg B disparate-impact + proxy detection | disparate_impact_test, protected_class_proxy_detection, decisioning_variable_audit |
| Customer Advocate | Adverse-action specificity + plain-language disclosure | adverse_action_codes, disclosure_quality_check, plain_language_score |
| Macro Contrarian | Through-cycle macro + sector cycle phase | macro_regime_classifier, sector_cycle_phase, recession_proxies |
- Node.js ≥ 20
- npm workspaces (bundled with Node)
- A modern desktop browser (for the WebXR scenes in fallback mode) or a Meta Quest 3S device on the same network
git clone https://github.com/alex-jb/embodied-compliance-council.git
cd embodied-compliance-council
npm installnpm run dev:tradingOpen http://localhost:5173. Drag to look around, press SPACE to fire a synthetic council decision, or click Enable hand tracking to drive verdicts with thumbs-up (approve), fist (block), or thumb-index pinch (escalate) via MediaPipe Hands.
npm run dev:bankingOpen http://localhost:5174. Banking maps the gestures to approve / decline / escalate.
npm run test # all workspacesCurrent coverage:
packages/spatial-gating-protocol/: 4/4 hash-chain tests (genesis, valid chain, payload tamper, link tamper).packages/hand-gestures/: 6/6 gesture classifier tests (thumbs-up, fist, pinch, open palm, empty, partial).
| Phase | What | Status |
|---|---|---|
| Phase 1 | Council-voice system prompts + tool restrictions | ✅ shipped 2026-06-16 |
| Phase 1 calibration | Brier + ECE + temperature + Reflexion + Kelly + Sakana DG | ✅ shipped in capstone-orallexa-calibration (sibling repo, 64 tests green) |
| Phase 2 | WebXR scenes + spatial gating protocol + hash-chain audit | ✅ shipped 2026-06-17 |
| Phase 2.5 | MediaPipe Hands gesture input (desktop webcam + Quest 3S XRHand) |
✅ webcam path shipped 2026-06-17; Quest path next |
| Phase 3 | Wire synthetic decisions to real Anthropic calls per voice | next |
| Phase B | Banking risk math integrated from Loredana Levitchi's research base | pending Loredana code share |
| Phase 4 | EU AI Act Article 14 + ECOA / Reg B compliance evaluation studies | Year 2 |
This codebase backs four academic deliverables:
- Capstone thesis (COM 6000-CP1, Prof. Michael Yang) — "Calibration Discipline as the Audit Primitive for Spatial Pre-Trade Compliance in Live Trading." Trading vertical, defended August 2026.
- Dr. NGO Final Project — banking-side embodied compliance for loan origination + wealth advisory, ECOA / Reg B framing.
- ICAIF 2027 paper — calibration discipline + multi-voice methodology for the trading vertical. Submission deadline ~2026-09-15.
- IEEE banking paper (joint with Loredana C. Levitchi) — disparate-impact detection via multi-voice fair-lending review. Target journal: IEEE TIFS or IEEE Computer.
Strategic and TAM analysis underlying the academic defense lives in docs/startup/ — these documents are research armor for the defense, not a startup execution plan.
council-diff— TypeScript / npm package implementing the 5-voice deliberation pattern as a standalone library (consumed by Slack and CLI surfaces).capstone-orallexa-calibration— Python calibration toolkit (ECE, temperature scaling, Reflexion, Kelly sizing, Sakana DG self-modification).memory-wall-tracker— Druckenmiller AI inference basket with Brier-audited daily research.
- Alex Xiaoyu Ji — primary developer (Katz School of Science and Health, Yeshiva University, Graduate Researcher)
- Prof. Michael Yang — capstone advisor, trading vertical
- Dr. NGO — banking vertical co-PI
- Loredana C. Levitchi — banking risk-math collaborator, IEEE paper co-author
Pull requests welcome but please open an issue first to discuss scope — this is an active research codebase, not a stable library.
MIT. See LICENSE.